New opportunity

AI Architect

Avahi · Banglore, India

About this role

Avahi seeks an AI Architect for its AI & Cloud practice. The role leads the design and implementation of secure, cost-efficient, production-grade AI systems for clients, with a focus on Amazon Bedrock, Bedrock Agents and Amazon OpenSearch Service. The posting lists Banglore (Remote), full-time employment and compensation of 50 LPA.

Responsibilities

Architect end-to-end AI/ML solutions using Bedrock foundation models. Design multi-agent systems with action groups, knowledge bases and guardrails; build embedding pipelines, vector indexes and retrieval patterns for RAG applications; and design semantic, lexical, k-NN vector and hybrid search.
Produce architecture blueprints, architecture decision records and reference designs. Build scalable AWS infrastructure, connect Bedrock workflows to AWS data services and event-driven pipelines, and apply VPC isolation, PrivateLink, KMS encryption and fine-grained IAM policies. Use infrastructure-as-code tools such as AWS CDK, CloudFormation or Terraform.
Advise clients on technical roadmaps, lead AI proofs of concept and present findings to engineering and executive stakeholders. Mentor teams on prompt engineering, model evaluation, LLM fine-tuning and responsible AI; conduct architecture reviews against AWS Well-Architected principles.
Collaborate with product managers, data scientists and software engineers. Track latency, accuracy, retrieval relevance, cost per query and hallucination rates, and deliver technical documentation, runbooks and client-facing materials.

Required qualifications

At least 8 years of software engineering or cloud architecture experience, including at least 3 years focused on AI/ML or generative AI solutions.
Hands-on Amazon Bedrock expertise, including model invocation APIs, Agents, Knowledge Bases and Guardrails; experience operating OpenSearch clusters and building RAG pipelines with enterprise data; strong Python proficiency.
Understanding of AWS networking, identity and security for AI workloads, plus experience with multi-account AWS architectures and enterprise-scale deployments.

Preferred qualifications

AWS Certified Solutions Architect certification is strongly preferred; AWS Certified Machine Learning – Specialty and familiarity with LangChain, LlamaIndex or comparable orchestration frameworks are pluses. Avahi describes a remote-first environment and offers mentorship and career-development opportunities.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

AI architectureGenerative AIMachine learningAmazon BedrockAmazon Bedrock AgentsAmazon Bedrock Knowledge BasesAmazon Bedrock GuardrailsAmazon OpenSearch ServiceRetrieval-augmented generationMulti-agent orchestrationPrompt engineeringModel evaluationLLM fine-tuningEmbedding pipelinesVector searchSemantic searchHybrid searchRelevance tuningPythonAWS LambdaAmazon ECSAmazon EKSAmazon API GatewayAWS Step FunctionsAmazon S3AWS IAMAWS GlueAmazon AthenaAmazon RDSAmazon DynamoDB

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Bedrock solution architecture
The role leads end-to-end AI/ML design using Bedrock foundation models and designs multi-agent systems with action groups, Knowledge Bases, and Guardrails. Prepare to explain architecture choices and their rationale, not just service familiarity.
Enterprise RAG and OpenSearch
The posting calls for embedding pipelines, vector indexes, retrieval patterns, and semantic, lexical, k-NN, and hybrid search. It also requires experience operating OpenSearch clusters and tuning relevance for RAG over enterprise data.
Secure, scalable AWS delivery
The architect designs AWS infrastructure and integrations with data services and event-driven pipelines, applies VPC isolation, PrivateLink, KMS, and fine-grained IAM, and drives infrastructure as code. Prepare to explain how those decisions support enterprise-scale, multi-account deployments.
Client-facing technical leadership
The role advises clients on roadmaps, leads AI proofs of concept, presents findings to engineering and executive stakeholders, and produces blueprints, decision records, and reference designs. Prepare to communicate both technical decisions and their business implications.
Evaluation and operational readiness
The architect defines and tracks latency, accuracy, retrieval relevance, cost per query, and hallucination rates; conducts Well-Architected reviews; and produces runbooks. Prepare to explain how these measures inform design changes and production operations.

Likely challenges to think through

Balancing retrieval quality with speed and cost
Likely preparation hypothesis: richer embeddings, larger retrieval sets, or hybrid search may improve relevance while increasing latency and cost per query. Prepare a way to compare retrieval options against the posting’s KPIs and explain when a simpler approach is preferable.
Constraining agent behavior without losing utility
Likely preparation hypothesis: multi-agent action groups and access to enterprise data can make useful workflows harder to secure and evaluate. Prepare to discuss guardrails, least-privilege IAM, data isolation, and tests for unsafe actions or unsupported answers.
Turning a proof of concept into an operable client platform
Likely preparation hypothesis: a fast PoC may not address multi-account deployment, infrastructure as code, documentation, or ongoing ownership. Prepare to describe the trade-off between rapid validation and production readiness, including what an architecture decision record, review, and runbook would capture.

Practice your answers

Technical practice

How would you choose a Bedrock foundation model for a client AI application?

What to coverThe role involves end-to-end design with Bedrock foundation models and calls for secure, cost-efficient production systems. Suggested answer: compare task quality, latency, cost per query and operational constraints, then explain how you would validate the choice against the client's needs.

How would you design a multi-agent workflow with Bedrock Agents?

What to coverThe posting names action groups, knowledge bases and guardrails as parts of multi-agent orchestration. Suggested answer: explain each agent's responsibility, how actions and retrieved information flow between them, and where guardrails and failure handling belong.

How would you build a RAG pipeline over enterprise data?

What to coverThe role calls for embedding pipelines, vector indexing and retrieval patterns that connect foundation models to enterprise data. Suggested answer: walk through ingestion, updating indexes, retrieving context and checking whether that context supports the generated answer.

When would you use lexical, vector or hybrid search in OpenSearch?

What to coverThe posting specifically calls for semantic and lexical search, k-NN vector search, hybrid search, index management and relevance tuning. Suggested answer: compare their strengths for different query types and describe how you would diagnose poor retrieval relevance.

How would you structure the AWS infrastructure for a production Bedrock application?

What to coverThe role lists Lambda, ECS/EKS, API Gateway, Step Functions, S3 and IAM as relevant platform services. Suggested answer: choose only the components a proposed workload needs, explaining scaling, orchestration and cost trade-offs rather than treating the list as a required stack.

How would you connect a Bedrock workflow to data services and events?

What to coverThe posting names Glue, Athena, RDS and DynamoDB alongside Kinesis, SQS and EventBridge. Suggested answer: trace a plausible data flow, distinguish synchronous from event-driven steps, and explain how you would handle retries and data freshness.

How would you secure an enterprise AI workload across AWS accounts?

What to coverThe role calls for VPC isolation, PrivateLink, KMS encryption, fine-grained IAM and experience with multi-account architectures. Suggested answer: explain access boundaries and data protection from ingestion through model invocation and retrieval, including how you would review permissions.

How would you make an AI architecture reproducible and reviewable?

What to coverThe posting asks for infrastructure as code, architecture blueprints, decision records and reviews against AWS Well-Architected principles. Suggested answer: describe how CDK, CloudFormation or Terraform could capture the design, and what trade-offs and review findings an ADR should record.

How would you evaluate and improve a Bedrock-based AI solution?

What to coverThe role names prompt engineering, model evaluation, fine-tuning, responsible AI and KPIs for accuracy, hallucinations, latency and cost per query. Suggested answer: define a representative evaluation set and explain how you would distinguish prompt, retrieval and model issues before choosing an improvement.

How would you troubleshoot a production RAG application's rising latency and falling answer quality?

What to coverThe posting calls for tracking latency, retrieval relevance and hallucination rates, as well as delivering runbooks. Suggested answer: isolate retrieval, model invocation and infrastructure contributions; describe the measurements and runbook steps that would guide a safe fix.

Behavioural practice

When have you translated an ambiguous client need into a technical roadmap?

What to coverThe role calls for advising clients on AI/ML architecture and turning business needs into actionable roadmaps. Suggested practice: use a real situation, your actions to clarify priorities and trade-offs, and the resulting decision or outcome.

Tell me about a proof of concept whose findings you presented to different audiences.

What to coverThe posting calls for leading AI PoCs and presenting findings to engineering and executive stakeholders. Suggested practice: describe the situation, how you framed evidence and limitations for each audience, and what resulted.

Describe a time you coached a team through an AI engineering decision.

What to coverThe role includes mentoring on prompt engineering, model evaluation, fine-tuning and responsible AI. Suggested practice: give a real situation, the specific guidance you provided, and how the team's approach or result changed.

Tell me about a cross-functional disagreement during solution delivery.

What to coverThe posting calls for collaboration with product managers, data scientists and software engineers throughout development. Suggested practice: explain the situation, how you worked through competing priorities, and the outcome without overstating your individual contribution.

Describe an architecture review that led you to change a design.

What to coverThe role includes architecture reviews against AWS Well-Architected principles and producing decision records and documentation. Suggested practice: identify the situation, the trade-off or risk you surfaced, the action you took, and the resulting design or operational outcome.

Sharpen your knowledge

Topics to prepare

Bedrock agent architecture
Suggested exercise: Diagram how a Bedrock foundation model, Agents, action groups, Knowledge Bases and Guardrails would interact in a multi-agent workflow; explain where each component belongs. The posting explicitly calls for these components and end-to-end AI architecture.
RAG and OpenSearch retrieval design
Suggested exercise: Compare lexical, semantic k-NN and hybrid retrieval for a hypothetical enterprise document collection, including embedding, indexing and relevance-tuning decisions. The posting requires RAG design and deep OpenSearch experience.
AWS security and account boundaries
Suggested exercise: Review a hypothetical AI architecture for VPC isolation, PrivateLink, KMS encryption, fine-grained IAM and multi-account separation; identify access and data-flow risks. These controls and enterprise-scale AWS architectures are stated in the posting.
Production AI evaluation and trade-offs
Suggested exercise: Define how to measure latency, accuracy, retrieval relevance, cost per query and hallucination rates for a proposed system, then explain a trade-off between two designs. The posting names these KPIs and calls for secure, cost-efficient production systems.
Architecture review and client decisions
Suggested exercise: Outline an architecture decision record comparing two Bedrock-based approaches, then explain the choice to both engineers and executives using AWS Well-Architected principles. The role explicitly involves ADRs, client advice, presentations and architecture reviews.

Skills to revise

Python Bedrock integration
Suggested exercise: Write a small Python example that invokes a Bedrock model and handles errors; separately map the configuration needed for an Agent action group, Knowledge Base and Guardrail. Python proficiency and hands-on Bedrock expertise are required.
OpenSearch indexing and relevance tuning
Suggested exercise: With sample documents, define a vector index, run k-NN and lexical searches, combine them into a hybrid result set, and inspect which results are relevant. The posting specifically cites OpenSearch cluster operation, index management and relevance tuning.
Enterprise-data RAG pipeline
Suggested exercise: Prototype ingestion of non-sensitive sample documents, embedding and retrieval, followed by a Bedrock-grounded response; test a question the documents cannot answer. The posting requires building and deploying RAG pipelines with enterprise data.
AWS infrastructure as code
Suggested exercise: Draft a small AWS CDK, CloudFormation or Terraform deployment for an AI workflow, specifying IAM permissions, encryption and network boundaries. These IaC options and security controls are explicitly listed; a sample deployment is preparation, not a stated interview task.
Proof-of-concept delivery
Suggested exercise: Prepare a concise PoC package with an architecture diagram, decision record, KPI results, operational runbook and two versions of a findings summary for engineering and executive audiences. The posting calls for PoCs, documentation, runbooks and stakeholder presentations.

Your preparation checklist

Map your evidence to the requirements
The posting asks for 8+ years in engineering or cloud architecture, 3+ years in AI/ML, hands-on Bedrock and OpenSearch, Python, and multi-account AWS experience. Prepare a brief evidence matrix with specific examples you can substantiate; distinguish required experience from preferred certifications.
Prepare a production AI project walkthrough
Select a project you can discuss and outline the business need, your architecture decisions, implementation role, deployment, trade-offs, and measured results. Connect it to Avahi’s focus on secure, cost-efficient enterprise AI without implying you used services you did not use.
Sketch a Bedrock agent architecture
For practice, create a one-page design showing model invocation, Bedrock Agents, action groups, Knowledge Bases, Guardrails, and integration points. Annotate when an agent is warranted and how you would handle failures and access boundaries.
Work through an OpenSearch RAG design
Draft an enterprise-data ingestion and embedding flow, index strategy, and retrieval path. Be ready to explain k-NN, lexical and hybrid search, relevance tuning, and OpenSearch cluster operations, which the posting explicitly emphasizes.
Review AWS deployment and security choices
Prepare a blueprint or decision record covering suitable AWS compute and data services, event-driven integration, infrastructure as code, VPC isolation, PrivateLink, KMS, fine-grained IAM, and multi-account boundaries. Explain trade-offs rather than treating every service as mandatory.
Build an evaluation and operations plan
Outline how you would measure the posting’s stated KPIs—latency, accuracy, retrieval relevance, cost per query, and hallucination rates—and turn findings into an architecture review, runbook, and next-step recommendations.
Rehearse client leadership and prepare questions
Practice presenting a PoC outcome and technical roadmap in versions for engineers and executives. Ask Avahi how client needs are translated into PoC scope, how architecture reviews use AWS Well-Architected principles, and which AI metrics matter most in production.

Availability last checked: 5 Nov 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in India. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…